Video wird geladen...

Video konnte nicht geladen werden

Zur Startseite

Running DeepSeek R1 on my desk Uses EXO Labs with Thunderbolt 5 interconnect (80Gbps) to run the full (671B, 8-bit) DeepSeek R1 distributed across 2 M3 Ultra 512GB Mac Studios (1TB total Unified Memory). Runs at 11 tok/sec. Theoretical max is ~20 tok/sec.

alexocheema's profile picture

Alex Cheema

54,707 subscribers

992,987 Aufrufe • vor 1 Jahr •via X (Twitter)

GamingBildungWissenschaft & Technologie
LIVE

Anya Rossi• Live Now

Private livecam show

0 Kommentare

Keine Kommentare verfügbar

Kommentare vom Original-Post werden hier angezeigt

Ähnliche Videos

Running Kimi K2.5 on my desk. Runs at 24 tok/sec with 2 x 512GB M3 Ultra Mac Studios connected with Thunderbolt 5 (RDMA) using EXO Labs / MLX backend. Yes, it can run clawdbot.
0:18

Running Kimi K2.5 on my desk. Runs at 24 tok/sec with 2 x 512GB M3 Ultra Mac Studios connected with Thunderbolt 5 (RDMA) using EXO Labs / MLX backend. Yes, it can run clawdbot.

Alex Cheema

3,313,322 Aufrufe • vor 7 Monaten

EXO 1.0 at Apple NeurIPS booth. 4 x 512GB M3 Ultra Mac Studios clustered all-to-all with RDMA over Thunderbolt 5. Runs DeepSeek v3.2 (8-bit) at 25 tok/sec with tensor parallelism.
0:15

EXO 1.0 at Apple NeurIPS booth. 4 x 512GB M3 Ultra Mac Studios clustered all-to-all with RDMA over Thunderbolt 5. Runs DeepSeek v3.2 (8-bit) at 25 tok/sec with tensor parallelism.

EXO Labs

186,329 Aufrufe • vor 9 Monaten

M4 Mac Mini AI Cluster Uses EXO Labs with Thunderbolt 5 interconnect (80Gbps) to run LLMs distributed across 4 M4 Pro Mac Minis. The cluster is small (iPhone for reference). It’s running Nemotron 70B at 8 tok/sec and scales to Llama 405B (benchmarks soon).
0:19

M4 Mac Mini AI Cluster Uses EXO Labs with Thunderbolt 5 interconnect (80Gbps) to run LLMs distributed across 4 M4 Pro Mac Minis. The cluster is small (iPhone for reference). It’s running Nemotron 70B at 8 tok/sec and scales to Llama 405B (benchmarks soon).

Alex Cheema

3,517,980 Aufrufe • vor 1 Jahr

Running the full GLM 4.7 (8-bit) on 2 x 512GB M3 Ultra Mac Studios Runs at 19.8 tok/sec with EXO Labs MLX RDMA backend (h/t Awni Hannun) & tensor parallel
0:12

Running the full GLM 4.7 (8-bit) on 2 x 512GB M3 Ultra Mac Studios Runs at 19.8 tok/sec with EXO Labs MLX RDMA backend (h/t Awni Hannun) & tensor parallel

Alex Cheema

156,874 Aufrufe • vor 9 Monaten

Running DeepSeek-V3 on M4 Mac Mini AI Cluster 671B MoE model distributed across 8 M4 Pro 64GB Mac Minis. Apple Silicon with unified memory is a great fit for MoE.
0:18

Running DeepSeek-V3 on M4 Mac Mini AI Cluster 671B MoE model distributed across 8 M4 Pro 64GB Mac Minis. Apple Silicon with unified memory is a great fit for MoE.

EXO Labs

719,293 Aufrufe • vor 1 Jahr

DeepSeek R1 distilled to Qwen 1.5B easily runs on my iPhone 16 with MLX swift. Here's the 4-bit model reasoning entirely on device at almost 60 toks/sec:
0:19

DeepSeek R1 distilled to Qwen 1.5B easily runs on my iPhone 16 with MLX swift. Here's the 4-bit model reasoning entirely on device at almost 60 toks/sec:

Awni Hannun

1,139,934 Aufrufe • vor 1 Jahr

First steps for a specialized DeepSeek v4 Flash inference engine focused on inference quality / stability at different quantizations, with networked API that is batching capable. This is the 2 bit quants model running on my M3 Max 128GB.
0:12

First steps for a specialized DeepSeek v4 Flash inference engine focused on inference quality / stability at different quantizations, with networked API that is batching capable. This is the 2 bit quants model running on my M3 Max 128GB.

antirez

14,220 Aufrufe • vor 4 Monaten

Qwen3.8-Flash-Next is starting to feel like the local model Opus fans have been waiting for. Someone ran the NVFP4 176B-class Flash-Next on 2× DGX Sparks, and the results are wild. Real measured scaling → C1: 44.2 tok/s → C2: 64.6 tok/s → C4: 86.8 tok/s aggregate The per-stream speed drops with concurrency, but total throughput keeps climbing. Long-context behavior was even more impressive: → 5K: needle retrieved → 21K: needle retrieved → 84K: needle retrieved → 167K: needle retrieved → 262K: prefill succeeded, but the window was saturated That 167K retrieval test is the one I care about. Long agent runs are where models usually start losing the plot. Flash-Next didn’t. It also held up surprisingly well on physics-heavy reasoning, artifact generation, research workflows, evidence checking, and long-horizon planning. The personality is interesting too. DeepSeek V4 Flash feels like the dependable workhorse. GLM-5.2 feels like the problem-solving machine. Qwen3.8-Flash-Next feels more insightful. It has that rare ability to understand what you’re actually asking rather than just following the surface pattern. The main weakness I’ve noticed is instruction following. It can occasionally drift between prose turns where DeepSeek and GLM stay tighter. And this is why the 256GB M5 Ultra conversation gets interesting. If Apple can pair that huge unified-memory pool with enough bandwidth, this model class becomes genuinely practical for long-running local agents. We’re talking about frontier-class reasoning on hardware sitting on a desk.
0:18

Qwen3.8-Flash-Next is starting to feel like the local model Opus fans have been waiting for. Someone ran the NVFP4 176B-class Flash-Next on 2× DGX Sparks, and the results are wild. Real measured scaling → C1: 44.2 tok/s → C2: 64.6 tok/s → C4: 86.8 tok/s aggregate The per-stream speed drops with concurrency, but total throughput keeps climbing. Long-context behavior was even more impressive: → 5K: needle retrieved → 21K: needle retrieved → 84K: needle retrieved → 167K: needle retrieved → 262K: prefill succeeded, but the window was saturated That 167K retrieval test is the one I care about. Long agent runs are where models usually start losing the plot. Flash-Next didn’t. It also held up surprisingly well on physics-heavy reasoning, artifact generation, research workflows, evidence checking, and long-horizon planning. The personality is interesting too. DeepSeek V4 Flash feels like the dependable workhorse. GLM-5.2 feels like the problem-solving machine. Qwen3.8-Flash-Next feels more insightful. It has that rare ability to understand what you’re actually asking rather than just following the surface pattern. The main weakness I’ve noticed is instruction following. It can occasionally drift between prose turns where DeepSeek and GLM stay tighter. And this is why the 256GB M5 Ultra conversation gets interesting. If Apple can pair that huge unified-memory pool with enough bandwidth, this model class becomes genuinely practical for long-running local agents. We’re talking about frontier-class reasoning on hardware sitting on a desk.

FHILY👑

20,253 Aufrufe • vor 27 Tagen

I can't believe this is real I have GLM 5.2 running 100% locally on my Mac Studio. 2 bit quant. The results I'm getting are better than Opus 4.8 It's now powering my Hermes Agent and Codex. 100% free, local, private super intelligence on my desk I also have it in a loop coding for me 24/7 now I thought we were at least a year away from this type of event. It happened today. The model takes up about 250gb of memory. So you can technically run it on a Mac Studio with 256gb, but you probably want the 512gb memory version (please tell me you listened to me 5 months ago when these were sitting on store shelves) With Fable gone, I now have Opus 4.8 level intelligence on my desk for free. This is the future. Local, private, secure, personal super intelligence. If you're still writing off local AI as a fad or engagement bait, you are officially delusional
0:18

I can't believe this is real I have GLM 5.2 running 100% locally on my Mac Studio. 2 bit quant. The results I'm getting are better than Opus 4.8 It's now powering my Hermes Agent and Codex. 100% free, local, private super intelligence on my desk I also have it in a loop coding for me 24/7 now I thought we were at least a year away from this type of event. It happened today. The model takes up about 250gb of memory. So you can technically run it on a Mac Studio with 256gb, but you probably want the 512gb memory version (please tell me you listened to me 5 months ago when these were sitting on store shelves) With Fable gone, I now have Opus 4.8 level intelligence on my desk for free. This is the future. Local, private, secure, personal super intelligence. If you're still writing off local AI as a fad or engagement bait, you are officially delusional

Alex Finn

623,903 Aufrufe • vor 3 Monaten

you can run claude code inside antigravity completely Free with zero credit card and no rate limits 😳 use openrouter’s free models + antigravity. no anthropic bill. no paid api keys. takes 10 minutes to set up. what you get during this setup: - full claude code agent experience - strong coding models (including deepseek-r1, qwen2.5-coder, llama-4, grok-4 free tier) - antigravity’s clean workspace and sandbox - unlimited usage (as long as you stay on free models) - easy model swapping - zero cost full setup guide (100% free): step 1: install antigravity -go to and install it -create a new workspace step 2: install claude code - inside antigravity, install the claude code extension from the marketplace - open the built-in terminal step 3: create openrouter free account -go to - sign up with google (no card needed) - go to keys and create a new api key step 4: set the environment variables -in antigravity terminal run: export ANTHROPIC_API_KEY=sk-or-xxx export OPENROUTER_API_KEY=sk-or-xxx step 5: launch claude code with free model -run this command: claude-code --model deepseek/deepseek-r1:free or try: qwen/qwen2.5-coder:free if you already have antigravity? skip straight to step 2. after 10 minutes you’ll have a full agentic coding setup running for free. this is currently one of the cheapest ways to run serious coding agents in 2026. bookmark this before they limit the free models.
0:17

you can run claude code inside antigravity completely Free with zero credit card and no rate limits 😳 use openrouter’s free models + antigravity. no anthropic bill. no paid api keys. takes 10 minutes to set up. what you get during this setup: - full claude code agent experience - strong coding models (including deepseek-r1, qwen2.5-coder, llama-4, grok-4 free tier) - antigravity’s clean workspace and sandbox - unlimited usage (as long as you stay on free models) - easy model swapping - zero cost full setup guide (100% free): step 1: install antigravity -go to and install it -create a new workspace step 2: install claude code - inside antigravity, install the claude code extension from the marketplace - open the built-in terminal step 3: create openrouter free account -go to - sign up with google (no card needed) - go to keys and create a new api key step 4: set the environment variables -in antigravity terminal run: export ANTHROPIC_API_KEY=sk-or-xxx export OPENROUTER_API_KEY=sk-or-xxx step 5: launch claude code with free model -run this command: claude-code --model deepseek/deepseek-r1:free or try: qwen/qwen2.5-coder:free if you already have antigravity? skip straight to step 2. after 10 minutes you’ll have a full agentic coding setup running for free. this is currently one of the cheapest ways to run serious coding agents in 2026. bookmark this before they limit the free models.

painn

32,057 Aufrufe • vor 3 Monaten

We are in an insane run of open-weight drops. Every modality, open source is winning. This is what an open source AI summer ☀️ looks like: 🧠 LLMs & Reasoning → DeepSeek-V4-Flash-0731 (my king 👑): 304B MoE refresh, Terminal-Bench 2.1 jumps 61.8→82.7 over the preview, DeepSWE 7.3→54.4. Closes in on Opus-4.8 on Agents' Last Exam (25.2 vs 25.7). MIT. → Muse-Glimmer-30B, from Meta (they are back!!): their first open agentic model. ~29.6B dense + perception encoder, 131k+ context, built to run fully local, no cloud. Apache 2.0. → Liquid AI LFM2.5-2.6B: 2.69B params, 131k context, 220 tok/s on an M5 Max in under 2.5GB RAM. Competitive with models 4x larger on agentic tasks. → inclusionAI Ling-3.0-flash: 124B total, only 5.1B active, ~12% the size of their old 1T flagship Ring-2.6, matches it on key benchmarks. MIT. → inclusionAI Ling-3.0-tiny: 7.9B total, 1.3B active, 86-90 tok/s on an M4 Pro MacBook at ~8GB peak memory. MIT. → NVIDIA Nemotron-3.5-Lightning-30B-A3B: hybrid Mamba-2+MoE+Attention, up to 1M context, runs on a single H100 or DGX Spark, SWE-bench Verified 52.8. → deepgrove maple-preview: 20B-A1B ternary-weight reasoner, 218 tok/s on a Mac mini M4, 5.3GB checkpoint. MIT. → BigBang-v1 (endless-frontier): fine-tuned from Qwen3.6-35B-A3B via a self-evolving generator/critic synthetic-data loop. Lands aggregate performance between DeepSeek V4 Flash (284B) and V4 Pro (1.6T), at 35B. Apache 2.0. 🎬 Video → MiniMax-H3: 33B dense omni model, native stereo audio, up to 2K/15s. 3.6k+ likes already. → Minimax-H3-Turbo (lightx2v): Apache-2.0 turbo distillation of H3 for fast inference. → Lightricks LTX-2.5: image-to-video update, custom Gemma-4-12B text encoder, a markedly stronger distilled model. 🔊 Voice → NVIDIA NemotronLabs VoiceChat-11B: full-duplex speech-to-speech, ~450ms turn-taking, #2 on open VoiceBench, and the first open full-duplex model with live tool-calling mid-conversation. 🛡️ Safety → Mistral Shieldstral-1.0-3B: 3B multimodal guardrail that takes your safety policy as plain text instead of fixed categories. Beats LlamaGuard-4-12B and ShieldGemma-9B on HarmBench (99.4) and ToxicChat (84.1) at a fraction of the size. Apache 2.0.
0:14

We are in an insane run of open-weight drops. Every modality, open source is winning. This is what an open source AI summer ☀️ looks like: 🧠 LLMs & Reasoning → DeepSeek-V4-Flash-0731 (my king 👑): 304B MoE refresh, Terminal-Bench 2.1 jumps 61.8→82.7 over the preview, DeepSWE 7.3→54.4. Closes in on Opus-4.8 on Agents' Last Exam (25.2 vs 25.7). MIT. → Muse-Glimmer-30B, from Meta (they are back!!): their first open agentic model. ~29.6B dense + perception encoder, 131k+ context, built to run fully local, no cloud. Apache 2.0. → Liquid AI LFM2.5-2.6B: 2.69B params, 131k context, 220 tok/s on an M5 Max in under 2.5GB RAM. Competitive with models 4x larger on agentic tasks. → inclusionAI Ling-3.0-flash: 124B total, only 5.1B active, ~12% the size of their old 1T flagship Ring-2.6, matches it on key benchmarks. MIT. → inclusionAI Ling-3.0-tiny: 7.9B total, 1.3B active, 86-90 tok/s on an M4 Pro MacBook at ~8GB peak memory. MIT. → NVIDIA Nemotron-3.5-Lightning-30B-A3B: hybrid Mamba-2+MoE+Attention, up to 1M context, runs on a single H100 or DGX Spark, SWE-bench Verified 52.8. → deepgrove maple-preview: 20B-A1B ternary-weight reasoner, 218 tok/s on a Mac mini M4, 5.3GB checkpoint. MIT. → BigBang-v1 (endless-frontier): fine-tuned from Qwen3.6-35B-A3B via a self-evolving generator/critic synthetic-data loop. Lands aggregate performance between DeepSeek V4 Flash (284B) and V4 Pro (1.6T), at 35B. Apache 2.0. 🎬 Video → MiniMax-H3: 33B dense omni model, native stereo audio, up to 2K/15s. 3.6k+ likes already. → Minimax-H3-Turbo (lightx2v): Apache-2.0 turbo distillation of H3 for fast inference. → Lightricks LTX-2.5: image-to-video update, custom Gemma-4-12B text encoder, a markedly stronger distilled model. 🔊 Voice → NVIDIA NemotronLabs VoiceChat-11B: full-duplex speech-to-speech, ~450ms turn-taking, #2 on open VoiceBench, and the first open full-duplex model with live tool-calling mid-conversation. 🛡️ Safety → Mistral Shieldstral-1.0-3B: 3B multimodal guardrail that takes your safety policy as plain text instead of fixed categories. Beats LlamaGuard-4-12B and ShieldGemma-9B on HarmBench (99.4) and ToxicChat (84.1) at a fraction of the size. Apache 2.0.

Victor M

55,281 Aufrufe • vor 1 Monat

I watched so you did not have to and here is the breakdown!!!! Nearly 2 hours. Plenty of nostalgia. Plenty of vibes. Lots of feelings. But when the net was wide open for actual clarification? Dead air. Total time Chris Albert spoke 1,501 seconds Which is 25 min and 1 sec. So on an almost 2 hour show, he spoke for about 25 minutes total. Roughly 21% of the entire show The remaining time was largely filled by the hosts themselves frequently cutting him off, reframing questions, finishing answers for him, or moving the conversation along before he ever had to sit in discomfort or explain anything fully. 0 minutes of that time addressed the major unresolved timeline and movement issues Oddly missing from this fact-based discussion No explanation for how Colin got home 20 min AFTER Chris that night, despite a four-minute head start No explanation for why Chris’s wife that went to 34 Fairview with donuts for brian Jr. pulled behind a car that was not there only an hour earlier with no explanation then, turned around, came back to get him, returned again, stayed 30 min, then left, all on the morning John was found. No mention of parents allegedly coordinating to quietly clean up a house-destroying party involving minors in which his son was one No discussion of repeated allegations of Colin being removed from bars for underage drinking. We were told this channel is about facts, not opinion. So we treated it like a data exercise. We timed everything Chris talked about, so you didn’t have to. Here’s how nearly two hours were spent 2 min saying they tried to “let the court system play out,” while explaining his New Year’s Day Twitter Space tour and condemning Sean McCabe’s comments 37 sec on moving to Canton in 1974 35 sec explaining it’s normal to work in the town you grew up in (we agree) 35 sec on Canton no longer being blue-collar and now very expensive 44 sec reminiscing about red paint stains in his childhood garage was misrepresented by the defense 1 min on town politics 15 sec explaining what a Select Board is 50 sec on handling public comment and being a “First Amendment guy… to a degree” 38 sec announcing he’s not running again 25 sec on how he keeps his cool during public comment 15 sec on pineapple pizza 52 sec explaining Colin was 17 when John died and can’t play college football with his cousin 45 sec expressing loss for how his family has been treated 10 sec referencing a blogger saying “horrible things” about Colin 26 sec blaming the media for selling and sensationalizing a story 20 sec saying he’s mind-boggled 48 sec explaining why the term “McAlberts” is dehumanizing 28 sec criticizing Yannetti’s performance was upsetting to him while stating John was his friend 42 sec explaining the fence-straddling photo was a joke 1:50 on “Nebbercracker” and a childhood yard incident 10 sec about his wife bringing donuts to a nephew for his birthday was not weird 32 sec blaming Karen for making his family look like villains 1 min on selling his house and convincing his brother to sell 24 sec saying Nebbercracker was the best part of the trial 1:12 on knowing Michael Proctor but not really knowing him 10 sec saying he was never alone with Proctor 11 sec correcting the “pin it on the girl” claim (which we agree with) 1 min praising Jillian Daniels as a great family supporter and defender 1 min explaining early lack of support made him sad 15 sec thanking the hosts 31 sec discussing starting a “Defend the Truth” media page 30 sec saying both sides were horrible and doxxed people 24 sec on feeling isolated at first 33 sec minimizing the Proctor gift as “just gratitude for solving a murder in one day” 28 sec saying he’s still friends with the O’Keefes and they eat at his pizza shop 38 sec saying Mr. O’Keefe sits with him for long periods of time 43 sec lamenting being made into villains 30 sec saying the case took years off his parents’ lives 35 sec thanking supporters and saying “the truth is on our side”
0:15

I watched so you did not have to and here is the breakdown!!!! Nearly 2 hours. Plenty of nostalgia. Plenty of vibes. Lots of feelings. But when the net was wide open for actual clarification? Dead air. Total time Chris Albert spoke 1,501 seconds Which is 25 min and 1 sec. So on an almost 2 hour show, he spoke for about 25 minutes total. Roughly 21% of the entire show The remaining time was largely filled by the hosts themselves frequently cutting him off, reframing questions, finishing answers for him, or moving the conversation along before he ever had to sit in discomfort or explain anything fully. 0 minutes of that time addressed the major unresolved timeline and movement issues Oddly missing from this fact-based discussion No explanation for how Colin got home 20 min AFTER Chris that night, despite a four-minute head start No explanation for why Chris’s wife that went to 34 Fairview with donuts for brian Jr. pulled behind a car that was not there only an hour earlier with no explanation then, turned around, came back to get him, returned again, stayed 30 min, then left, all on the morning John was found. No mention of parents allegedly coordinating to quietly clean up a house-destroying party involving minors in which his son was one No discussion of repeated allegations of Colin being removed from bars for underage drinking. We were told this channel is about facts, not opinion. So we treated it like a data exercise. We timed everything Chris talked about, so you didn’t have to. Here’s how nearly two hours were spent 2 min saying they tried to “let the court system play out,” while explaining his New Year’s Day Twitter Space tour and condemning Sean McCabe’s comments 37 sec on moving to Canton in 1974 35 sec explaining it’s normal to work in the town you grew up in (we agree) 35 sec on Canton no longer being blue-collar and now very expensive 44 sec reminiscing about red paint stains in his childhood garage was misrepresented by the defense 1 min on town politics 15 sec explaining what a Select Board is 50 sec on handling public comment and being a “First Amendment guy… to a degree” 38 sec announcing he’s not running again 25 sec on how he keeps his cool during public comment 15 sec on pineapple pizza 52 sec explaining Colin was 17 when John died and can’t play college football with his cousin 45 sec expressing loss for how his family has been treated 10 sec referencing a blogger saying “horrible things” about Colin 26 sec blaming the media for selling and sensationalizing a story 20 sec saying he’s mind-boggled 48 sec explaining why the term “McAlberts” is dehumanizing 28 sec criticizing Yannetti’s performance was upsetting to him while stating John was his friend 42 sec explaining the fence-straddling photo was a joke 1:50 on “Nebbercracker” and a childhood yard incident 10 sec about his wife bringing donuts to a nephew for his birthday was not weird 32 sec blaming Karen for making his family look like villains 1 min on selling his house and convincing his brother to sell 24 sec saying Nebbercracker was the best part of the trial 1:12 on knowing Michael Proctor but not really knowing him 10 sec saying he was never alone with Proctor 11 sec correcting the “pin it on the girl” claim (which we agree with) 1 min praising Jillian Daniels as a great family supporter and defender 1 min explaining early lack of support made him sad 15 sec thanking the hosts 31 sec discussing starting a “Defend the Truth” media page 30 sec saying both sides were horrible and doxxed people 24 sec on feeling isolated at first 33 sec minimizing the Proctor gift as “just gratitude for solving a murder in one day” 28 sec saying he’s still friends with the O’Keefes and they eat at his pizza shop 38 sec saying Mr. O’Keefe sits with him for long periods of time 43 sec lamenting being made into villains 30 sec saying the case took years off his parents’ lives 35 sec thanking supporters and saying “the truth is on our side”

Dixie Normus

17,063 Aufrufe • vor 8 Monaten

Qwen3.8-Flash-Next now reaches ~43 tok/s after a 122,902-token prompt on ONE DGX Spark. ⚡🚀 MTP k=2 won my draft-depth sweep, with +42.5% mean decode over no draft. The PLE table stays fully on-device. I promised the deeper MTP tests. Here are the results, and now you can explore them in an interactive benchmark page too. 𝗧𝗪𝗢 𝗗𝗥𝗔𝗙𝗧 𝗧𝗢𝗞𝗘𝗡𝗦 𝗪𝗢𝗡 Mean single-request decode with 32K context configured: MTP k=2: 39.21 tok/s MTP k=3: 36.42 tok/s MTP k=1: 35.18 tok/s No draft: 27.51 tok/s k=2 also produced the fastest individual sweep run: 41.34 tok/s. Four runs each for no draft, k=1 and k=2. Seven for k=3. Decode excludes time to first token. Here, k means speculative draft depth, not quantization bits. k=3 produced more tokens per step, but the extra drafting work did not pay off in throughput. k=2 is my current pick for this setup. 𝗧𝗛𝗘 𝟭𝟮𝟯𝗞-𝗧𝗢𝗞𝗘𝗡 𝗣𝗥𝗢𝗠𝗣𝗧 𝗧𝗘𝗦𝗧 I then ran a separate long-prompt comparison: Actual input: 122,902 tokens Configured context: 262,144 Requested output: 128 tokens One request at a time MTP k=2: ~43 tok/s No draft: 26.2 tok/s Time to first token: 110.6 seconds with MTP 107.0 seconds without it The win here is generation speed, not faster prefill. To keep the scope clear: 256K was the configured limit. This was a real ~123K input, not a completely filled 256K window or a full k sweep at that depth. 𝗣𝗟𝗘 𝗦𝗧𝗔𝗬𝗦 𝗢𝗡 𝗧𝗛𝗘 𝗦𝗣𝗔𝗥𝗞 Whole model on-device: 78.57 GiB Packed 5-bit PLE table: 30.4 GiB, included in that total No NVMe PLE offload in this build. This is still turboderp’s 3.05bpw_h5_ng5 EXL3 pack, served through my vllm-exl3 integration. My work here is the serving integration and testing. These are preliminary performance measurements, not a quality evaluation or a claim of bit-exact full-output parity. 𝗘𝗫𝗣𝗟𝗢𝗥𝗘 𝗧𝗛𝗘 𝗥𝗘𝗦𝗨𝗟𝗧𝗦 The benchmark page has the individual sweep values, long-prompt comparison, and measurement scope. You can play the animation, export the charts, or download the HTML and data to render them yourself. No Spark needed to view the results. Credit to turboderp / ExLlamaV3 for the pack and kernels, vLLM for the serving engine, and Qwen Qwen Developers for the model. Recipe + reproduction: Interactive benchmark:
0:19

Qwen3.8-Flash-Next now reaches ~43 tok/s after a 122,902-token prompt on ONE DGX Spark. ⚡🚀 MTP k=2 won my draft-depth sweep, with +42.5% mean decode over no draft. The PLE table stays fully on-device. I promised the deeper MTP tests. Here are the results, and now you can explore them in an interactive benchmark page too. 𝗧𝗪𝗢 𝗗𝗥𝗔𝗙𝗧 𝗧𝗢𝗞𝗘𝗡𝗦 𝗪𝗢𝗡 Mean single-request decode with 32K context configured: MTP k=2: 39.21 tok/s MTP k=3: 36.42 tok/s MTP k=1: 35.18 tok/s No draft: 27.51 tok/s k=2 also produced the fastest individual sweep run: 41.34 tok/s. Four runs each for no draft, k=1 and k=2. Seven for k=3. Decode excludes time to first token. Here, k means speculative draft depth, not quantization bits. k=3 produced more tokens per step, but the extra drafting work did not pay off in throughput. k=2 is my current pick for this setup. 𝗧𝗛𝗘 𝟭𝟮𝟯𝗞-𝗧𝗢𝗞𝗘𝗡 𝗣𝗥𝗢𝗠𝗣𝗧 𝗧𝗘𝗦𝗧 I then ran a separate long-prompt comparison: Actual input: 122,902 tokens Configured context: 262,144 Requested output: 128 tokens One request at a time MTP k=2: ~43 tok/s No draft: 26.2 tok/s Time to first token: 110.6 seconds with MTP 107.0 seconds without it The win here is generation speed, not faster prefill. To keep the scope clear: 256K was the configured limit. This was a real ~123K input, not a completely filled 256K window or a full k sweep at that depth. 𝗣𝗟𝗘 𝗦𝗧𝗔𝗬𝗦 𝗢𝗡 𝗧𝗛𝗘 𝗦𝗣𝗔𝗥𝗞 Whole model on-device: 78.57 GiB Packed 5-bit PLE table: 30.4 GiB, included in that total No NVMe PLE offload in this build. This is still turboderp’s 3.05bpw_h5_ng5 EXL3 pack, served through my vllm-exl3 integration. My work here is the serving integration and testing. These are preliminary performance measurements, not a quality evaluation or a claim of bit-exact full-output parity. 𝗘𝗫𝗣𝗟𝗢𝗥𝗘 𝗧𝗛𝗘 𝗥𝗘𝗦𝗨𝗟𝗧𝗦 The benchmark page has the individual sweep values, long-prompt comparison, and measurement scope. You can play the animation, export the charts, or download the HTML and data to render them yourself. No Spark needed to view the results. Credit to turboderp / ExLlamaV3 for the pack and kernels, vLLM for the serving engine, and Qwen Qwen Developers for the model. Recipe + reproduction: Interactive benchmark:

Cruz

12,258 Aufrufe • vor 16 Tagen

This guy built a mini AI farm out of 4 Nvidia boxes It does not look like a data center. It looks like a stack of small machines sitting next to a laptop. But each box is a DGX Spark with Grace Blackwell inside, 128GB unified memory, and enough room to run models normal gaming GPUs cannot even open. Using the launch price from the article, 4 of them is almost $12,000 of local AI compute on one desk. That sounds expensive until you compare it to cloud GPUs. A serious AI builder can burn $1,500 to $3,000 a month renting A100s and H100s for client work, fine-tunes, agents and 70B models. He basically moved that bill from the cloud into hardware he owns. 4 Nvidia boxes. 512GB unified memory. No hourly meter running in the background. No rented GPUs eating the margin every time an agent runs too long. The funny part is most people still think local AI means a slow laptop running a toy model. Meanwhile guys like this are stacking compute at home. Save this, local AI is turning into the new mining farm.
0:11

This guy built a mini AI farm out of 4 Nvidia boxes It does not look like a data center. It looks like a stack of small machines sitting next to a laptop. But each box is a DGX Spark with Grace Blackwell inside, 128GB unified memory, and enough room to run models normal gaming GPUs cannot even open. Using the launch price from the article, 4 of them is almost $12,000 of local AI compute on one desk. That sounds expensive until you compare it to cloud GPUs. A serious AI builder can burn $1,500 to $3,000 a month renting A100s and H100s for client work, fine-tunes, agents and 70B models. He basically moved that bill from the cloud into hardware he owns. 4 Nvidia boxes. 512GB unified memory. No hourly meter running in the background. No rented GPUs eating the margin every time an agent runs too long. The funny part is most people still think local AI means a slow laptop running a toy model. Meanwhile guys like this are stacking compute at home. Save this, local AI is turning into the new mining farm.

Gipp 🦅

591,405 Aufrufe • vor 3 Monaten

This is the most hilarious thing I saw and did today Ran gemma-4-12B-coder-fable5-composer2.5-v1-GGUF locally with 8 GB VRAM at 20+ tok/sec Anthropic's Claude Fable 5 launched June 9. By June 12 it was banned. I can't access it. You can't either. But here's the twist: I'm running a model trained on its chain of thought at 20 tok/s on my RTX 4060 8GB. Locally. Offline. No cloud. No export control. Enter: Gemma4-12B-Coder GGUF (Q4_K_M) Base: Google's gemma-4-12B-it Fine-tuned on verifiable Python CoT data: - Primary: Composer 2.5 real reasoning traces (only passing solutions kept) - Auxiliary: Fable 5 used to redo the hard cases Composer missed. Every training example's reasoning led to code that actually ran. No hallucinated logic. Llama.cpp flags: -m gemma4-coding-Q4_K_M.gguf -cnv -ngl 44 -c 64000 -v (huggingface model link in comments) Flag breakdown: -ngl 44 → offload 44 layers to GPU (tune this for your VRAM) -c 64000 → 64K context window -cnv → conversation/chat mode -v → verbose output The irony writes itself. Anthropic spent weeks telling the world Fable 5 (mythos) is too powerful to release. Then released it. Then got banned from serving it, including their own researchers. Meanwhile: a Gemma 4 12B fine tune, trained on Fable 5's reasoning, runs fully offline on my mid range consumer GPU No API. No cloud. Just me and llama.cpp. This is why local AI matters. Check out the model's link in the comments. How's your experience been with this model?
0:15

This is the most hilarious thing I saw and did today Ran gemma-4-12B-coder-fable5-composer2.5-v1-GGUF locally with 8 GB VRAM at 20+ tok/sec Anthropic's Claude Fable 5 launched June 9. By June 12 it was banned. I can't access it. You can't either. But here's the twist: I'm running a model trained on its chain of thought at 20 tok/s on my RTX 4060 8GB. Locally. Offline. No cloud. No export control. Enter: Gemma4-12B-Coder GGUF (Q4_K_M) Base: Google's gemma-4-12B-it Fine-tuned on verifiable Python CoT data: - Primary: Composer 2.5 real reasoning traces (only passing solutions kept) - Auxiliary: Fable 5 used to redo the hard cases Composer missed. Every training example's reasoning led to code that actually ran. No hallucinated logic. Llama.cpp flags: -m gemma4-coding-Q4_K_M.gguf -cnv -ngl 44 -c 64000 -v (huggingface model link in comments) Flag breakdown: -ngl 44 → offload 44 layers to GPU (tune this for your VRAM) -c 64000 → 64K context window -cnv → conversation/chat mode -v → verbose output The irony writes itself. Anthropic spent weeks telling the world Fable 5 (mythos) is too powerful to release. Then released it. Then got banned from serving it, including their own researchers. Meanwhile: a Gemma 4 12B fine tune, trained on Fable 5's reasoning, runs fully offline on my mid range consumer GPU No API. No cloud. Just me and llama.cpp. This is why local AI matters. Check out the model's link in the comments. How's your experience been with this model?

Alok

575,300 Aufrufe • vor 3 Monaten

THAT $70 "RUN YOUR OWN LLMS" PI KIT CAN'T RUN A SINGLE LLM. IT'S A VISION CHIP WITH NO RAM. that clip sells a raspberry pi 5 in a slick case with an ai accelerator and the caption "your own llms." clean build, fun kit. the claim is where it breaks. the fine print: the popular $70 pi ai kit uses a hailo-8l, 13 tops. it's built for vision, object detection and image processing, and it has no memory of its own. so it cannot run large language models. full stop the board that actually can is a different one: the newer ai hat+ 2, hailo-10h, 40 tops, with 8gb of dedicated ram. that's $130, not $70 and even that runs only tiny models. llama 3.2 at 1b, qwen 2.5 at 1.5b, deepseek r1 at 1.5b. edge llms live in the 1-7b range, against cloud models at 500b to 2 trillion so the honest pitch: for $130 you can run a very small language model on a pi, slowly, as a fun learning project. that's real and it's cool. "your own llms" on a $70 vision kit is not. why this keeps happening: "ai kit" and a big "tops" number sell. tops sounds like intelligence. but tops measures vision-style math, not whether the chip has the memory to hold a language model. the spec that matters for llms is ram, and the cheap kit has none. the honest caveats, both ways: the $70 kit is genuinely great, just at vision. cameras, object detection, that's its job the $130 hat really does run small llms locally, which a pi couldn't do at all two years ago. that's progress "small" is the load-bearing word. don't expect gpt at home on a pi the takeaway: before you buy a kit because the caption says llm, check two numbers. not the tops. the ram, and the size of the model it can actually load. no 70-dollar miracle, no gpt in a pi case, no tops number that means what you think. save this before you buy the wrong kit for the word on the box.
0:17

THAT $70 "RUN YOUR OWN LLMS" PI KIT CAN'T RUN A SINGLE LLM. IT'S A VISION CHIP WITH NO RAM. that clip sells a raspberry pi 5 in a slick case with an ai accelerator and the caption "your own llms." clean build, fun kit. the claim is where it breaks. the fine print: the popular $70 pi ai kit uses a hailo-8l, 13 tops. it's built for vision, object detection and image processing, and it has no memory of its own. so it cannot run large language models. full stop the board that actually can is a different one: the newer ai hat+ 2, hailo-10h, 40 tops, with 8gb of dedicated ram. that's $130, not $70 and even that runs only tiny models. llama 3.2 at 1b, qwen 2.5 at 1.5b, deepseek r1 at 1.5b. edge llms live in the 1-7b range, against cloud models at 500b to 2 trillion so the honest pitch: for $130 you can run a very small language model on a pi, slowly, as a fun learning project. that's real and it's cool. "your own llms" on a $70 vision kit is not. why this keeps happening: "ai kit" and a big "tops" number sell. tops sounds like intelligence. but tops measures vision-style math, not whether the chip has the memory to hold a language model. the spec that matters for llms is ram, and the cheap kit has none. the honest caveats, both ways: the $70 kit is genuinely great, just at vision. cameras, object detection, that's its job the $130 hat really does run small llms locally, which a pi couldn't do at all two years ago. that's progress "small" is the load-bearing word. don't expect gpt at home on a pi the takeaway: before you buy a kit because the caption says llm, check two numbers. not the tops. the ram, and the size of the model it can actually load. no 70-dollar miracle, no gpt in a pi case, no tops number that means what you think. save this before you buy the wrong kit for the word on the box.

RetroChainer

11,100 Aufrufe • vor 2 Monaten

vllm-exl3 v0.3.0 is LIVE with custom native CUDA kernels for 2-bit EXL3 on NVIDIA DGX Spark GB10. GLM-5.3-Flash-EXL3-K2 jumped from 16.9 → 24.6 tok/s average single-stream decode, a +45.6% gain. Coding hit 27.6 tok/s, +85.6%. 🚀 The previous ExLlamaV3-backed path inside vLLM was leaving a lot of GB10 bandwidth on the table. So I rewrote the hot path specifically for EXL3 on Blackwell sm_121: → in-register Trellis dequantization → native fused MoE decode → power-of-two chunked prefill GEMM → parallel NVMe pre-warm Then I tested it side-by-side on physical DGX Spark hardware using my GLM-5.3-Flash-EXL3-K2 pack and live vLLM HTTP streaming. 🚀 𝗗𝗘𝗖𝗢𝗗𝗘 𝗧𝗛𝗥𝗢𝗨𝗚𝗛𝗣𝗨𝗧 Single-stream C1: Coding 14.9 → 27.6 tok/s +85.6% Prose 13.7 → 24.6 tok/s +79.3% Reasoning 18.9 → 25.1 tok/s +32.7% Summary 17.1 → 25.6 tok/s +50.0% Format 16.3 → 24.0 tok/s +47.7% Average: 16.9 → 24.6 tok/s 𝗡𝗘𝗧 𝗚𝗔𝗜𝗡: +45.6% ⏱️ 𝗙𝗜𝗥𝗦𝗧-𝗧𝗢𝗞𝗘𝗡 𝗥𝗘𝗦𝗣𝗢𝗡𝗦𝗜𝗩𝗘𝗡𝗘𝗦𝗦 Coding TTFT: 2,344 ms → 859 ms That is a 63.3% reduction, or about 2.7× faster to first token. Follow-up turn with prefix cache hit: 5,608 ms → 3,588 ms 1.56× faster. ⚡ 𝗪𝗛𝗔𝗧 𝗖𝗛𝗔𝗡𝗚𝗘𝗗 𝗢𝗡 𝗧𝗛𝗘 𝗚𝗣𝗨 40 routed-MoE layers: 19.9 ms → 11.5 ms per token Per-layer MoE compute: 497 μs → 287.8 μs That removes 8.4 ms of MoE compute from every generated token. Total per-step wall time: 59.2 ms → 40.6 ms -31.4% The key is `p2b_fused_moe`. Instead of expanding EXL3 weights through a traditional intermediate path, the new kernel performs Trellis dequantization in-register while executing the routed expert computation. The weights stay compressed until the GPU actually needs them. 🔥 𝗣𝗥𝗘𝗙𝗜𝗟𝗟 𝗚𝗢𝗧 𝗔 𝗡𝗔𝗧𝗜𝗩𝗘 𝗣𝗔𝗧𝗛 𝗧𝗢𝗢 The new `exl3_gemm` uses power-of-two chunked prefill GEMM. Measured: 7.85 TFLOPS 13.0× faster than the legacy prefill kernel 1,875 tok/s cold prefill sustained across 65K context 💾 𝗧𝗛𝗘 𝗕𝗢𝗢𝗧 𝗣𝗔𝗧𝗛 𝗡𝗘𝗘𝗗𝗘𝗗 𝗪𝗢𝗥𝗞 𝗧𝗢𝗢 Loading a ~91 GiB model is part of the user experience. Standard shard loading is mostly serial. The updated recipe parallelizes NVMe pre-warm across 8 workers so the storage controller gets used properly instead of feeding a ~100 GiB model one shard at a time. That turns boot-time storage into another optimization target instead of something we simply accept. 💡 𝗧𝗪𝗢 𝗦𝗘𝗥𝗩𝗜𝗡𝗚 𝗙𝗟𝗔𝗚𝗦 𝗪𝗢𝗥𝗧𝗛 𝗞𝗡𝗢𝗪𝗜𝗡𝗚 `--long-prefill-token-threshold 1024` Prevents giant prefill chunks from monopolizing step budgets and starving parallel decode sessions. `--enable-prefix-caching` Avoids paying for the same conversational prefix again on follow-up turns. 📦 𝗘𝗩𝗘𝗥𝗬𝗧𝗛𝗜𝗡𝗚 𝗜𝗦 𝗢𝗣𝗘𝗡 vllm-exl3: GLM-5.3-Flash one-Spark recipe: Model: This is why I like working at the kernel level. The model did not change. The quant did not change. The hardware did not change. The execution path did. 16.9 → 24.6 tok/s. 🛠️ vLLM turboderp
0:11

vllm-exl3 v0.3.0 is LIVE with custom native CUDA kernels for 2-bit EXL3 on NVIDIA DGX Spark GB10. GLM-5.3-Flash-EXL3-K2 jumped from 16.9 → 24.6 tok/s average single-stream decode, a +45.6% gain. Coding hit 27.6 tok/s, +85.6%. 🚀 The previous ExLlamaV3-backed path inside vLLM was leaving a lot of GB10 bandwidth on the table. So I rewrote the hot path specifically for EXL3 on Blackwell sm_121: → in-register Trellis dequantization → native fused MoE decode → power-of-two chunked prefill GEMM → parallel NVMe pre-warm Then I tested it side-by-side on physical DGX Spark hardware using my GLM-5.3-Flash-EXL3-K2 pack and live vLLM HTTP streaming. 🚀 𝗗𝗘𝗖𝗢𝗗𝗘 𝗧𝗛𝗥𝗢𝗨𝗚𝗛𝗣𝗨𝗧 Single-stream C1: Coding 14.9 → 27.6 tok/s +85.6% Prose 13.7 → 24.6 tok/s +79.3% Reasoning 18.9 → 25.1 tok/s +32.7% Summary 17.1 → 25.6 tok/s +50.0% Format 16.3 → 24.0 tok/s +47.7% Average: 16.9 → 24.6 tok/s 𝗡𝗘𝗧 𝗚𝗔𝗜𝗡: +45.6% ⏱️ 𝗙𝗜𝗥𝗦𝗧-𝗧𝗢𝗞𝗘𝗡 𝗥𝗘𝗦𝗣𝗢𝗡𝗦𝗜𝗩𝗘𝗡𝗘𝗦𝗦 Coding TTFT: 2,344 ms → 859 ms That is a 63.3% reduction, or about 2.7× faster to first token. Follow-up turn with prefix cache hit: 5,608 ms → 3,588 ms 1.56× faster. ⚡ 𝗪𝗛𝗔𝗧 𝗖𝗛𝗔𝗡𝗚𝗘𝗗 𝗢𝗡 𝗧𝗛𝗘 𝗚𝗣𝗨 40 routed-MoE layers: 19.9 ms → 11.5 ms per token Per-layer MoE compute: 497 μs → 287.8 μs That removes 8.4 ms of MoE compute from every generated token. Total per-step wall time: 59.2 ms → 40.6 ms -31.4% The key is `p2b_fused_moe`. Instead of expanding EXL3 weights through a traditional intermediate path, the new kernel performs Trellis dequantization in-register while executing the routed expert computation. The weights stay compressed until the GPU actually needs them. 🔥 𝗣𝗥𝗘𝗙𝗜𝗟𝗟 𝗚𝗢𝗧 𝗔 𝗡𝗔𝗧𝗜𝗩𝗘 𝗣𝗔𝗧𝗛 𝗧𝗢𝗢 The new `exl3_gemm` uses power-of-two chunked prefill GEMM. Measured: 7.85 TFLOPS 13.0× faster than the legacy prefill kernel 1,875 tok/s cold prefill sustained across 65K context 💾 𝗧𝗛𝗘 𝗕𝗢𝗢𝗧 𝗣𝗔𝗧𝗛 𝗡𝗘𝗘𝗗𝗘𝗗 𝗪𝗢𝗥𝗞 𝗧𝗢𝗢 Loading a ~91 GiB model is part of the user experience. Standard shard loading is mostly serial. The updated recipe parallelizes NVMe pre-warm across 8 workers so the storage controller gets used properly instead of feeding a ~100 GiB model one shard at a time. That turns boot-time storage into another optimization target instead of something we simply accept. 💡 𝗧𝗪𝗢 𝗦𝗘𝗥𝗩𝗜𝗡𝗚 𝗙𝗟𝗔𝗚𝗦 𝗪𝗢𝗥𝗧𝗛 𝗞𝗡𝗢𝗪𝗜𝗡𝗚 `--long-prefill-token-threshold 1024` Prevents giant prefill chunks from monopolizing step budgets and starving parallel decode sessions. `--enable-prefix-caching` Avoids paying for the same conversational prefix again on follow-up turns. 📦 𝗘𝗩𝗘𝗥𝗬𝗧𝗛𝗜𝗡𝗚 𝗜𝗦 𝗢𝗣𝗘𝗡 vllm-exl3: GLM-5.3-Flash one-Spark recipe: Model: This is why I like working at the kernel level. The model did not change. The quant did not change. The hardware did not change. The execution path did. 16.9 → 24.6 tok/s. 🛠️ vLLM turboderp

Cruz

21,173 Aufrufe • vor 20 Tagen

you're paying $20/mo for something your $500 GPU can already do. Gemma 4 26B A4B QAT MoE + Hermes Agent running on a single RTX 4060 (8GB VRAM). Built a vision capable, 100% free, 100% local, private AI assistant that lives in my Chrome browser. No API keys. No cloud. No subscriptions. 100% vibe coded. 0% handholding. It has full context of whatever's on my screen can answer questions, summarize pages, extract data, and see images. Same local model handles everything, no external calls, ever. keep reading for the model and hermes agent tips i learnt while building this locally. Here's the exact setup for anyone running local LLMs on 6-8 GB VRAM: llama.cpp server flags (on my NVIDIA RTX 4060 8gb VRAM): -m gemma-4-26B-A4B-it-qat-UD-Q4_K_XL.gguf --cache-type-k q8_0 --cache-type-v q8_0 -c 150000 --port 8080 Throughput with quantization: Prefill: 200-250 tokens/sec Decode: 20-25 tokens/sec reduce context if oom on 6 gb vram card. Key learnings: - Quantize KV cache to q8 for faster prefill/decode. Prefill goes from 100-150 (unquantized) to 200-250 tok/s (q8). - But watch out, once actual context grows past ~50k tokens on high entropy workloads, q8 KV quantization can cause hallucinations. Low entropy workloads are mostly unaffected. If you see it happening, drop the quantization. This is common across all local models. - In Hermes Agent settings -> Memory & Context, bump compression threshold from default 0.5 to 0.7. Default triggers way too frequent context compression and eats time. Up next: add persistent memory, web search, tool calling, streaming output and whatever you suggest. Running a 26B MoE with vision + 150k context window on 8GB VRAM would've sounded impossible 6 months ago. Works the same on the NVIDIA RTX 3060 Ti, 3070, 4060 Ti, 5060, 2080, or any 8GB card. VRAM is the only requirement. Local AI agents are closer than people think. You just need to know where the knobs are. Model's Unsloth quant hugging face link in the comments. Have you tried Hermes agent by Nous Research yet? What are you building with local LLMs? Drop it below, let's see what this community is shipping.
0:19

you're paying $20/mo for something your $500 GPU can already do. Gemma 4 26B A4B QAT MoE + Hermes Agent running on a single RTX 4060 (8GB VRAM). Built a vision capable, 100% free, 100% local, private AI assistant that lives in my Chrome browser. No API keys. No cloud. No subscriptions. 100% vibe coded. 0% handholding. It has full context of whatever's on my screen can answer questions, summarize pages, extract data, and see images. Same local model handles everything, no external calls, ever. keep reading for the model and hermes agent tips i learnt while building this locally. Here's the exact setup for anyone running local LLMs on 6-8 GB VRAM: llama.cpp server flags (on my NVIDIA RTX 4060 8gb VRAM): -m gemma-4-26B-A4B-it-qat-UD-Q4_K_XL.gguf --cache-type-k q8_0 --cache-type-v q8_0 -c 150000 --port 8080 Throughput with quantization: Prefill: 200-250 tokens/sec Decode: 20-25 tokens/sec reduce context if oom on 6 gb vram card. Key learnings: - Quantize KV cache to q8 for faster prefill/decode. Prefill goes from 100-150 (unquantized) to 200-250 tok/s (q8). - But watch out, once actual context grows past ~50k tokens on high entropy workloads, q8 KV quantization can cause hallucinations. Low entropy workloads are mostly unaffected. If you see it happening, drop the quantization. This is common across all local models. - In Hermes Agent settings -> Memory & Context, bump compression threshold from default 0.5 to 0.7. Default triggers way too frequent context compression and eats time. Up next: add persistent memory, web search, tool calling, streaming output and whatever you suggest. Running a 26B MoE with vision + 150k context window on 8GB VRAM would've sounded impossible 6 months ago. Works the same on the NVIDIA RTX 3060 Ti, 3070, 4060 Ti, 5060, 2080, or any 8GB card. VRAM is the only requirement. Local AI agents are closer than people think. You just need to know where the knobs are. Model's Unsloth quant hugging face link in the comments. Have you tried Hermes agent by Nous Research yet? What are you building with local LLMs? Drop it below, let's see what this community is shipping.

Alok

36,691 Aufrufe • vor 2 Monaten

THREE 3090s ON ONE BOARD GIVE YOU 72GB OF VRAM AND KILL YOUR $200 CLAUDE CODE AND $200 OPENAI BILL people are pulling three used 3090s off ebay for around $2,100 total and stacking them in one tower to build a dedicated ai rig. that pools 72gb of vram for less than what a single rtx 5090 retails for alibaba shipped qwen 3.6 27b in april under apache 2.0. on realworldqa vision it scores 84.1 against claude 4.5 opus at 77.0. on ifbench instructions it lands at 76.5 against claude's 58.0 a single 3090 already runs qwen 3.6 27b with eight gigs of headroom. three of them in parallel handle larger models like deepseek r1 70b and qwen 235b without breaking a sweat a heavy ai user pays $200 claude code, $200 chatgpt pro plus $40 cursor and gemini. that's $5,280 a year and the rig pays itself off before month nine on $8 a month in electricity setup is one shell command for ollama, one to pull the model, one environment variable to point claude code at localhost. cli stays identical, nothing leaves the network, requests stop costing money bookmark this and read the article below
0:16

THREE 3090s ON ONE BOARD GIVE YOU 72GB OF VRAM AND KILL YOUR $200 CLAUDE CODE AND $200 OPENAI BILL people are pulling three used 3090s off ebay for around $2,100 total and stacking them in one tower to build a dedicated ai rig. that pools 72gb of vram for less than what a single rtx 5090 retails for alibaba shipped qwen 3.6 27b in april under apache 2.0. on realworldqa vision it scores 84.1 against claude 4.5 opus at 77.0. on ifbench instructions it lands at 76.5 against claude's 58.0 a single 3090 already runs qwen 3.6 27b with eight gigs of headroom. three of them in parallel handle larger models like deepseek r1 70b and qwen 235b without breaking a sweat a heavy ai user pays $200 claude code, $200 chatgpt pro plus $40 cursor and gemini. that's $5,280 a year and the rig pays itself off before month nine on $8 a month in electricity setup is one shell command for ollama, one to pull the model, one environment variable to point claude code at localhost. cli stays identical, nothing leaves the network, requests stop costing money bookmark this and read the article below

starmex

16,719 Aufrufe • vor 3 Monaten

Run Gemma 4 26B MoE on 8GB VRAM with 250k context at 20+ tokens/sec If you own any 8GB VRAM graphics card, stop what you are doing. Local AI just had its absolute "Holy Shit" moment for budget hardware. Yesterday, I benchmarked Unsloth Gemma 4 12B Q4_K_XL on an 8GB card. The community went wild but immediately demanded more: "Can we run a 25B+ model on budget GPUs?" Today, I’m delivering exactly that. I am running a massive 26B parameter Mixture of Experts (MoE) model locally on a standard 8GB VRAM setup with 250k full native context!. If you own an RTX 3060, 3070, 4060, or any budget GPU with 8GB of VRAM, the local AI paradigm has completely changed. The performance metrics are astonishing: - 20 tokens/sec flat decode throughput. - Stable, flat decode speed even with massive prompts. - I threw a 60k token prompt at it, and it still clocked in at 20 TPS without dropping a single frame. # What about prefill? Yes, Time To First Token (TTFT) is slightly high when swallowing massive contexts. But with a solid 200 tokens/sec prefill speed, the wait is barely noticeable and highly usable. And this is running completely without Multi Token Prediction (MTP) active. How is this possible? It’s the magic of Google's new QAT (Quantization Aware Training) quants for Gemma 4. The model weight file (unsloth gemma-4-26B-A4B-it-qat-UD-Q4_K_XL.gguf) is only 13.2 GB, making it the ultimate local powerhouse. # The Test Setup: CPU: Intel Core i7 RAM: 16GB System RAM GPU: NVIDIA GeForce RTX 4060 Laptop GPU (8GB VRAM) # The Secret Sauce (The -cmoe Flag) To make this work properly on any 8GB card, you must use the -cmoe (CPU MoE) flag in llama.cpp. This flag isolates the heavy MoE expert weights directly to system memory (CPU/RAM) while letting your GPU focus strictly on the Attention layers and the KV Cache. It prevents VRAM spillage and holds the throughput rock solid. # The flags: -m "gemma-4-26B-A4B-it-qat-UD-Q4_K_XL.gguf" -cmoe -c 248000 -v Once running, just open the UI on localhost and toggle the new reasoning lightbulb icon in the text input box to watch the model perform multi step thinking. Are you still running smaller models, or are you ready to scale up your budget local setups? Let's discuss in the replies
0:19

Run Gemma 4 26B MoE on 8GB VRAM with 250k context at 20+ tokens/sec If you own any 8GB VRAM graphics card, stop what you are doing. Local AI just had its absolute "Holy Shit" moment for budget hardware. Yesterday, I benchmarked Unsloth Gemma 4 12B Q4_K_XL on an 8GB card. The community went wild but immediately demanded more: "Can we run a 25B+ model on budget GPUs?" Today, I’m delivering exactly that. I am running a massive 26B parameter Mixture of Experts (MoE) model locally on a standard 8GB VRAM setup with 250k full native context!. If you own an RTX 3060, 3070, 4060, or any budget GPU with 8GB of VRAM, the local AI paradigm has completely changed. The performance metrics are astonishing: - 20 tokens/sec flat decode throughput. - Stable, flat decode speed even with massive prompts. - I threw a 60k token prompt at it, and it still clocked in at 20 TPS without dropping a single frame. # What about prefill? Yes, Time To First Token (TTFT) is slightly high when swallowing massive contexts. But with a solid 200 tokens/sec prefill speed, the wait is barely noticeable and highly usable. And this is running completely without Multi Token Prediction (MTP) active. How is this possible? It’s the magic of Google's new QAT (Quantization Aware Training) quants for Gemma 4. The model weight file (unsloth gemma-4-26B-A4B-it-qat-UD-Q4_K_XL.gguf) is only 13.2 GB, making it the ultimate local powerhouse. # The Test Setup: CPU: Intel Core i7 RAM: 16GB System RAM GPU: NVIDIA GeForce RTX 4060 Laptop GPU (8GB VRAM) # The Secret Sauce (The -cmoe Flag) To make this work properly on any 8GB card, you must use the -cmoe (CPU MoE) flag in llama.cpp. This flag isolates the heavy MoE expert weights directly to system memory (CPU/RAM) while letting your GPU focus strictly on the Attention layers and the KV Cache. It prevents VRAM spillage and holds the throughput rock solid. # The flags: -m "gemma-4-26B-A4B-it-qat-UD-Q4_K_XL.gguf" -cmoe -c 248000 -v Once running, just open the UI on localhost and toggle the new reasoning lightbulb icon in the text input box to watch the model perform multi step thinking. Are you still running smaller models, or are you ready to scale up your budget local setups? Let's discuss in the replies

Alok

292,770 Aufrufe • vor 3 Monaten